fix: tolerate typo'd render-tag names in echo strip; warn on degenerate end_turn turns

[ranxianglei]

Fixes

#673 — follow-up to #644/#646 (v0.1.97 incident family): a model-emitted render-tag echo whose tag name is typo'd (an acp permutation plus one stray char, e.g. acpi/acip) bypassed every exact-spelling pattern in the tag-echo filter. The 43-char fragment then became the turn's entire visible text: stop_reason=end_turn, zero tool_use, complete next-step plan sitting in the thinking block — the agent stalls mid-orchestration until manually nudged.

What changed

Gap 1 — tolerant name matching (src/loop/tag-echo-filter.ts)

• All strip patterns (PAIRED / LONE_OPEN / LONE_CLOSE / TRUNC_* / PARTIAL_TAIL / RENDER_TAG_DETECT) now match the acplike name set: permutations of acp plus one inserted a/c/p/i char (66 members, longest-first), built programmatically so the set stays auditable.
• Every occurrence still requires a strict name boundary (whitespace or close after the name), so legit text is untouched — verified: C includes of acpi/acpi.h, <caption>, <app id="1">, uppercase ACPI_DEVICE, underscore triggers, a < b.
• The swallow-until-close scan now finds loose close variants (mixed correct-open / typo-close echoes from the later dumps).
• Existing caps unchanged (hold 128 / definite drop 4096 / swallow release 80); worst-case false-positive surface bounded by them.
• New lifetime stats() (inputChars/outputChars/dropped) so callers can do turn-level accounting across intermediate flushes.
• src/loop/adapter-responses.ts module-level RENDER_TAG_RE loosened to the same name set.

Gap 2 — degenerate-turn detection (WARN only)

• New src/degenerate-turn.ts: pure predicate + message builder. Fires when a terminal turn ends with zero visible text AND zero tool calls (thinking may be non-empty — that is exactly the incident shape).
• Wired into all three compress-loop adapters (anthropic → end_turn, openai → stop, responses → completed) and both plugin passthrough pipes (chat SSE covers OpenAI + Anthropic wires; Responses pipe). In the pipes, visible-text accounting counts only content/text fields — thinking/reasoning never counts, matching the client-visible definition.
• Deliberately no auto-retry: retry interplay with billing, kernel state, and client-side retry loops needs owner sign-off. The warn gives operators a grep-able signal ([degenerate-turn]) immediately; auto-retry can land behind it as a follow-up.

Tests

• Split-parity harness extended with typo/mixed/safe cases (every split position x seeds).
• Gate engagement for typo'd openers/closers; safe-word survival static+streaming parity.
• stats() accounting; degenerateTurnWarning matrix.
• Adapter e2e: thinking-only + echo-only anthropic turn warns; clean turn does not.
• Pipe level: strips typo'd tags on the wire; warns on degenerate turns; no warn when visible text exists or a tool call is present.

Pre-flight

• npm run typecheck clean
• npm test: 1279/1279 pass
• npm run build clean (dist self-contained)
• E2E suite not run here (needs a live Responses upstream); unit coverage exercises every changed path

[github-actions]

📦 Built Package Artifact

Branch: 2026-09-10_acip-tag-echo-tolerance (da1ab44)

Option A — Install from npm PR tag (recommended)

npm install -g billion-context@pr-674

Each push to this PR publishes a new version under the pr-674 npm tag.

Option B — Download artifact

1. Download the artifact from the Actions run
2. Extract the tarball and install:

tar xzf billion-context-pr674.tgz
npm install -g package

---

This comment is automatically updated on each push.